You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The data of the past five years is collected from Thomson Reuters (now called Refinitiv). We denote the time horizon before 2017-04-25 as training period. We use that period to fit a regression model. From 2017-04-25 to the end of our dataset is defined as testing period. We use our model from training horizon to predict "reasonable" range of the value of NOK in testing horizon. The regression result comes out as below. We have a pretty high R squared. All T stats and F stats seem to be significant. I have to take back my words. Brent shows a major influence on NOK which implies Norway still heavily depends on petrolchemical industry. As the summary suggests, there could be multicollinearity (condition number is large and R squared is large). Obviously, Brent Crude and US Dollar should be negatively correlated. Most commodity future contracts are priced in US dollar. When US dollar appreciates or depreciates, the underlying commodity price is likely to go the opposite direction. There could be a cointegration relationship between Sterling and Euro for pre-Brexit time as well.
33
+
The data of the past five years is collected from Thomson Reuters (now called Refinitiv). We denote the time horizon before 2017-04-25 as training period. We use that period to fit a regression model. From 2017-04-25 to the end of our dataset is defined as testing period. We use our model from training horizon to predict "reasonable" range of the value of NOK in testing horizon. The regression result comes out as below. We have a pretty high R squared. All T stats and F stats seem to be significant. I have to take back my words. Brent shows a major influence on NOK which implies Norway still heavily depends on petrochemical industry. As the summary suggests, there could be multicollinearity (condition number is large and R squared is large). Obviously, Brent Crude and US Dollar should be negatively correlated. Most commodity future contracts are priced in US dollar. When US dollar appreciates or depreciates, the underlying commodity price is likely to go the opposite direction. There could be a cointegration relationship between Sterling and Euro for pre-Brexit time as well.
@@ -44,15 +44,15 @@ Before backtesting, we ought to set up thresholds for signal generation. One sig
44
44
45
45
To our surprise, there is a strong momentum after our model breaks. The momentum is independent of the selection of our training horizon. In another word, no matter which period we use as training horizon, the strong momentum always follows through right after the model breaks. The failure of this model doesn't really upset me. On the contrary, it is a blessing in disguise. What I see is an opportunity for trend following strategy! We will get back to that very soon.
46
46
47
-
What also interests me is why the model fails after 2017/11/15. Well, there is no way that a universal model can exist and work forever. Most models have short memory due to the dynamic environment of macroeconomics and geo-politics. Based on <a href=https://en.wikipedia.org/wiki/Adaptive_market_hypothesis>adpative market hypothesis</a> developed by Professor Andrew Lo from MIT, the players in the market are always evolving along with the fast-paced environment. Therefore, we should anticipate our model to lose its power sooner or later. Still, what is the cause of the break? For a quantamental analysis, we have only completed the quant part. I would not get into too much of rigorous and boring part of the fundamentals. Instead, I intend to bring up a short but insightful discussion here. What could possibly be the reason of this dramatic change? Could it be that Saudi and Iran endorsed an extension of oil production cap to boost up the oil price on that particular date? Or Donald Trump got elected as POTUS so he would encourage a weak US dollar and lift up restrictions on oil export as promised during his campaign? If we consider the price of NOK as a stochastic process, we can decompose NOK price into long term trend and short term random disturbance. Well, apparently short term is dominated by Brent Crude. It partially justifies our model. What really drives the long term trend though?
47
+
What also interests me is why the model fails after 2017/11/15. Well, there is no way that a universal model can exist and work forever. Most models have short memory due to the dynamic environment of macroeconomics and geo-politics. Based on <a href=https://en.wikipedia.org/wiki/Adaptive_market_hypothesis>adaptive market hypothesis</a> developed by Professor Andrew Lo from MIT, the players in the market are always evolving along with the fast-paced environment. Therefore, we should anticipate our model to lose its power sooner or later. Still, what is the cause of the break? For a quantamental analysis, we have only completed the quant part. I would not get into too much of rigorous and boring part of the fundamentals. Instead, I intend to bring up a short but insightful discussion here. What could possibly be the reason of this dramatic change? Could it be that Saudi and Iran endorsed an extension of oil production cap to boost up the oil price on that particular date? Or Donald Trump got elected as POTUS so he would encourage a weak US dollar and lift up restrictions on oil export as promised during his campaign? If we consider the price of NOK as a stochastic process, we can decompose NOK price into long term trend and short term random disturbance. Well, apparently short term is dominated by Brent Crude. It partially justifies our model. What really drives the long term trend though?
Ta-da, its Euro. As Norway is in EEA, its economic tie with EU totally dominates the long term trend of NOK. From the normalized figure, we clearly see the trend of both NOK and EUR are somewhat correlated. To get a formal conclusion, we need a cointegration test. Nevertheless, there is no Johansen Test in statsmodel package (not on the date when the first draft of oil money was finished, but there is now). We would have to use <a href=https://en.wikipedia.org/wiki/Cointegration#Engle–Granger_two-step_method>Engle-Granger two step test</a>. I have to honorably mention that this method is co-developed by the mentor of my mentor, Robert F. Engle, a Nobel Laureate! Unfortunately, we can't get any confirmation of cointegration from the test. The residual of the first step regression is not stationary under <a href=https://en.wikipedia.org/wiki/Augmented_Dickey–Fuller_test>Augmented Dickey-Fuller Test</a>. Sometimes we have to use the old fashion way to make a qualitative judgement, rule of thumb. Most researchers in big organizations recalibrate their forecast with policy or experience or consensus views or simply instinct (as an insider, I guarantee you this is 100% true). This is the moment that I exercise my sacred right to declare that EUR is the driver of NOK's long term trend!
Well, I am not an economist (even though my work has a lot to do with them). I do not expect to really find out what happened on 2017/11/15 through our quantamental analysis. Let's call it an end to this twist (if you are an economics-savvy person who is eager to know more, feel free to read <a href=https://www.norges-bank.no/globalassets/upload/publikasjoner/economic_bulletin/2000-04/factorsthat.pdf>this paper</a> by the Central Bank of Norway). In fact, we can simplify our model to NOK driven by Brent. From an econometrician's perspective, it does not sound like a good idea. Every coefficient of the model is statistically significant. Adding more variables does not worsen AIC, BIC or adjusted R squared. There is no incentive for us to remove variables as more information in the model is always better. From a trader's perspective, the requirement is a universal model. Ideally, the model contains two variables, one is the underlying petrocurrency, the other is the local crude oil contract. In this sense, the reduced form model can be replicated to any petrocurrency without too much focus on analyzing trade partners or exporting products. The argument here could be, is linear regression model too naïve? <a href=https://en.wikipedia.org/wiki/Nick_Patterson_(scientist)>Nick Patterson</a>, a cold war cryptographer who later worked in Renaissance Technologies, said (check <a href=http://www.thetalkingmachines.com/episodes/ai-safety-and-legacy-bletchley-park>here</a> for the orginial version of 45-minute-long interview),
55
+
Well, I am not an economist (even though my work has a lot to do with them). I do not expect to really find out what happened on 2017/11/15 through our quantamental analysis. Let's call it an end to this twist (if you are an economics-savvy person who is eager to know more, feel free to read <a href=https://www.norges-bank.no/globalassets/upload/publikasjoner/economic_bulletin/2000-04/factorsthat.pdf>this paper</a> by the Central Bank of Norway). In fact, we can simplify our model to NOK driven by Brent. From an econometrician's perspective, it does not sound like a good idea. Every coefficient of the model is statistically significant. Adding more variables does not worsen AIC, BIC or adjusted R squared. There is no incentive for us to remove variables as more information in the model is always better. From a trader's perspective, the requirement is a universal model. Ideally, the model contains two variables, one is the underlying petrocurrency, the other is the local crude oil contract. In this sense, the reduced form model can be replicated to any petrocurrency without too much focus on analyzing trade partners or exporting products. The argument here could be, is linear regression model too naïve? <a href=https://en.wikipedia.org/wiki/Nick_Patterson_(scientist)>Nick Patterson</a>, a cold war cryptographer who later worked in Renaissance Technologies, said (check <a href=http://www.thetalkingmachines.com/episodes/ai-safety-and-legacy-bletchley-park>here</a> for the original version of 45-minute-long interview),
56
56
57
57
`One tool that Renaissance uses is simple regression with one target and one independent.`
0 commit comments